graphify-agents

Graphify for Agents

Reading raw files into context is the dominant token cost of codebase-understanding work. Graphify builds a knowledge graph once (graphify-out/graph.json) and answers structural questions by traversing it — returning the relevant nodes and edges instead of whole files. This skill covers using that graph to make agent workflows cheaper and more targeted.

The project's published benchmark reports "71.5x fewer tokens per query vs reading raw files" on a mixed corpus, with smaller gains (≈5.4x on a 4-file corpus, ≈1x on a 6-file library) — the advantage scales with corpus size. Measure a specific repo with graphify benchmark before quoting a number; small repos see little benefit, so reserve the graph for large corpora.

For installing graphify and the full CLI surface, load the graphify skill.

The Core Pattern: Query, Don't Read

Question shape	Without graphify	With graphify
"How does auth connect to the request pipeline?"	Grep, open 6–10 files, read each	graphify query "how does auth connect to the request pipeline?" → relevant nodes + edges
"What is SwinTransformer and what touches it?"	Read the class + every caller	graphify explain "SwinTransformer"
"What breaks if I change add?"	Trace callers by hand	graphify affected "add"
"How do these two modules connect?"	Read both, infer	graphify path "DigestAuth" "Response"

query returns within a token budget (--budget, default 2000), so an agent gets a bounded, relevant slice rather than unbounded file contents.

Integration with /core:agent-loop

Graphify slots into the loop's existing phases without replacing them. Load /core:agent-loop for the phase model.

Phase 1 (Pre-flight / research). Before decomposition, build or refresh the graph, then have the research/Explore subagent query it instead of fanning out file reads:

mise run graphify:update      # AST-only refresh, no API key needed
graphify query "where is <epic-area> implemented and what does it depend on?"